annotating reads vs. peaks

[rebrendi]

Hi all,

I am asking a stupid question, but I am now really interested. Is there any reason, apart from decreasing the amount of computation time, for processing reads into clustered peaks and then annotating those peaks? Suppose you have a good machine and a lot of computer time. Then why would not you just annotate the primary reads and then make all statistics on them, without losing some valuable information on the intermediate step of peak identification?

Thanks!

[dpryan]

I would imagine that the non-uniformly distribute coverage of regions in the control dataset(s) used for comparison could make doing the statistics with raw reads rather more complicated.

[rebrendi]

I would imagine that the non-uniformly distribute coverage of regions in the control dataset(s) used for comparison could make doing the statistics with raw reads rather more complicated.

If reads are distributed non-uniformly, peaks would be also distributed non-uniformly, no?

[dpryan]

If reads are distributed non-uniformly, peaks would be also distributed non-uniformly, no?

Perhaps an example would be helpful. Suppose you have a number of regions of the genome that, for whatever reason, have a large number of reads aligning to them in both your experimental and control datasets and that the number of reads aligning to these regions is more or less the same across datasets (i.e., there's not really a peak here). Suppose also that you have a number of other regions with fewer raw reads, but they would be called as peaks due the different number of aligned reads in the two datasets. If you annotated all of the raw reads according to proximity to some feature (or inclusion of some feature, or pretty much anything else) and then ran statistics on that then the signal that you're interested in is likely to be drowned in the noise of those regions with much greater numbers of aligned reads. This is what I meant by the non-uniform distribution of reads in control datasets. I suppose you could consider peak calling as normalizing for read depth, or something like that.

There may be other reasons, but that's the one that's apparent to me. I'll add, it's also nice to have a list of sites with which to then do follow-up validation.

[rebrendi]

Perhaps an example would be helpful. Suppose you have a number of regions of the genome that, for whatever reason, have a large number of reads aligning to them in both your experimental and control datasets and that the number of reads aligning to these regions is more or less the same across datasets (i.e., there's not really a peak here). Suppose also that you have a number of other regions with fewer raw reads, but they would be called as peaks due the different number of aligned reads in the two datasets. If you annotated all of the raw reads according to proximity to some feature (or inclusion of some feature, or pretty much anything else) and then ran statistics on that then the signal that you're interested in is likely to be drowned in the noise of those regions with much greater numbers of aligned reads.

That is actually an example which has problems exactly because of using peaks instead of reads, and it could have been resolved in the case if individual reads, not peaks are being correlated.

[dpryan]

That is actually an example which has problems exactly because of using peaks instead of reads, and it could have been resolved in the case if individual reads, not peaks are being correlated.

You have that backwards. The regions with fewer overall reads but significant enrichment by group are the interesting regions. I would encourage you to put together a quick monte-carlo simulation if this isn't clear.